Back to Projects
Osteo-Vision — Android App - Image 1

Osteo-Vision — Android App

Machine Learning Engineer•2021
KotlinTensorFlow LiteXGBoostMaterial DesignNNAPI
View Source

Overview

An on-device osteoporosis detection app that processes X-ray images and form inputs using a hybrid deep learning model, with no cloud dependency or persistent storage.

Key Features

  • User inputs tabular health data via a form and uploads X-ray image
  • TFLite-quantized VGG-19 model extracts image features on-device
  • Combined prediction from CNN (image) and XGBoost (form data)
  • No database or cloud connection—results computed instantly
  • Material Design UI with multilingual support (EN, ES, FR)

Challenges & Solutions

Challenges

  • Running inference efficiently on mid-range Android hardware
  • Combining visual and structured input for a single prediction
  • Ensuring data privacy without storing any user data
  • Designing an intuitive, multilingual mobile experience

Solutions

  • Optimized TFLite model with NNAPI delegate for hardware acceleration
  • Feature fusion of CNN embeddings and form vectors for XGBoost
  • Processed everything on-device—no images or data ever stored
  • Used Android resource bundles and string localization for language support

Results

  • 92% diagnostic accuracy (ROC-AUC: 0.93 on 5K-image test set)
  • 60% lower inference time using NNAPI over CPU fallback
  • 100% offline-capable with zero cloud or database usage